Agentic search: an agent that plans, reads, refines and cites

Classic search answers one query at a time. You type, you get a list, and you decide what to do next. Agentic search hands that loop to an AI agent: it breaks a question into searches, reads what comes back, works out what is still missing, searches again and finally writes an answer with its sources. For operations and business teams the appeal is obvious, because the tedious part of research is exactly that loop. The risks are less obvious, and most of them come from the agent now making decisions you used to make yourself. This page explains how agentic search works, how it differs from one-shot search, and what a team should decide before an agent searches on its behalf.

One-shot search and agentic search side by side

Neither is better in general. If you know exactly what you are looking for, a single search is faster, cheaper and easier to check. Agentic search earns its cost when the question has several parts, when the right terms are not obvious at the start, or when the answer has to be assembled from many sources.

The loop, step by step

Most agentic search systems follow the same pattern, whatever name they go by:

1. Plan. The agent restates the question and splits it into sub-questions. “Which of our suppliers changed their payment terms this year?” becomes a list of suppliers, the places where such changes get announced, and a definition of what counts as a change.
2. Search. It issues queries for each sub-question, often with different phrasing, languages or filters.
3. Read. It opens the most promising results and pulls out the passages that bear on the question. Most of the cost and most of the value sit here: snippets alone are too thin to answer from.
4. Judge and refine. It compares what it has with what the plan needs. Gaps lead to new queries; contradictions lead to a search for a better source; new terms found inside documents lead to searches it could not have written at the start.
5. Stop. It ends when the sub-questions are covered, when further searching adds nothing new, or when a limit on time, steps or cost is reached.
6. Cite. It writes the answer, links each claim to the passage it came from, and should also list what it could not find.

The refine step is what separates agentic search from running a fixed list of queries in one go. The agent picks up vocabulary as it reads, which is why it can reach documents a prepared query list would never touch. Systems that run this loop for a long time on broad questions are usually sold as deep research; the mechanics are the same, only the scale differs.

What it is good at, and what it is not

Agentic search works well for multi-part questions, for unfamiliar fields where the vocabulary has to be learned along the way, for building lists from scattered sources, and for checking a claim against several independent pages.

It is a poor fit when the source of truth is already known, in which case query that system directly; when the answer must be exhaustive and provably so, because a search loop can never prove that nothing was missed; and when the question is really a judgement call that sources cannot settle. It also struggles where material sits behind logins or in systems the agent has no right to open. A link to a page is not permission to get past its login.

Where agentic search goes wrong

• Question drift. Each refinement shifts the question a little. After a dozen steps the agent may be answering a neighbour of what you asked.
• Stopping too early. The agent finds a plausible answer on the first page and declares the job done. Recall problems are invisible inside a fluent summary.
• Stopping too late. Without a limit, a curious agent keeps reading and the bill keeps growing.
• Weak sources. The agent treats a forum post, an outdated copy or a page written to influence AI answers as equal to a primary source.
• Citations that do not hold. A footnote pointing to a real page that does not say what the sentence claims. A generated link is not evidence until someone confirms it; citations that hold up goes into this.
• Silent gaps. Sub-questions with no answer vanish from the final text instead of being reported as open.

How to measure whether a search tool supports this loop well is covered in recall and accuracy in agent search.

What your team should decide before the agent searches

Most of these failures are prevented by decisions taken before the first query, not by checking afterwards. Write down five things:

• The question, fixed. If the question changes during the work, that is a new plan, not a quiet expansion of the old one. A changed question needs a new plan makes the case.
• The sources that may answer. An allowlist of domains or collections, or at least a list of excluded ones. See deciding which sources may answer a question.
• A budget and a stop rule. Maximum steps, time or spend, and what “covered” means for each sub-question.
• The evidence format. Every claim with a quoted passage, the page address and the retrieval date.
• What happens to gaps. Unanswered sub-questions are listed as open, not dropped.

In GoCux these decisions take the shape of a research plan: the question, its ordered steps and a source allowlist. The plan is recorded as a draft and has to be approved before anything runs, and the run cannot reach beyond what was approved; an allowlist left empty allows no source at all. Findings backed by a quotation are kept apart from questions no source answered, so a gap stays visible as an open question. GoCux does not search the open web itself. Its Knowledge search answers within one project's own documents, and web material comes in through an authorized browsing or export step, carrying its original page address and the date it was retrieved.

A brief you can reuse, and a review checklist

Give the agent a short written brief before it starts. A sample:

Question: Which of the twelve suppliers on the attached list announced a change to payment terms in the last twelve months? Sources: the suppliers' own websites and the official trade registry only. Stop when every supplier is marked changed, unchanged or not found, or after forty searches. For each change, quote the passage and give the page address and retrieval date. List suppliers with no source as not found.

When the answer comes back, review it with these questions:

• Does the answer address the question as written, rather than a neighbour of it?
• Did the agent stay inside the allowed sources?
• If you open three cited passages at random, can you find the claimed sentence in each?
• Are the unanswered parts listed as open rather than simply absent?
• Did the run stop for a stated reason: coverage, no new information, or the budget?

To see how search fits into longer, multi-step agent work, continue with agentic workflows.